rpotter6298
708fbc70ce
- Introduced `bridge_attention_ceiling_check.py` for variance decomposition analysis on bridge attention configurations. - Added `bridge_attention_readout.py` to perform per-tower gate and contribution readouts, including AUC sanity checks. - Created multiple JSON configuration files for backbone replication experiments, including anonymous CV variants and basic backbones. - Implemented sensitivity experiments to evaluate the impact of axial length inclusion and EfficientNetV2-M performance at higher resolutions. - Added a memory probe script to assess GPU memory usage during training with EfficientNetV2-M.
38 lines
1.6 KiB
JSON
38 lines
1.6 KiB
JSON
[
|
|
{
|
|
"_note": "refugelike (resnet50 + REFUGE fundus pretraining), freeze stem only (1/5 blocks). Tests whether even minimal anchoring helps stability. Bumped to 10 reps for the methods ablation table.",
|
|
"run_name": "experiments/freeze_sweep/refugelike_freeze20",
|
|
"reps": 10,
|
|
"tower_overrides": {
|
|
"img": { "args": { "backbone": "refugelike", "freeze_ratio": 0.2 } }
|
|
}
|
|
},
|
|
|
|
{
|
|
"_note": "Freeze stem + layer1 (2/5 blocks). Keeps low-level conv filters fixed, lets layers 2-4 + fc adapt. Bumped to 10 reps to confirm the 0.9105 result vs the 10-rep baseline_ensemble at 0.8958.",
|
|
"run_name": "experiments/freeze_sweep/refugelike_freeze40",
|
|
"reps": 10,
|
|
"tower_overrides": {
|
|
"img": { "args": { "backbone": "refugelike", "freeze_ratio": 0.4 } }
|
|
}
|
|
},
|
|
|
|
{
|
|
"_note": "Freeze stem + layer1 + layer2 (3/5 blocks). Only the deep semantic layers adapt — most aggressive practical setting before model loses capacity. Bumped to 10 reps for the methods ablation table.",
|
|
"run_name": "experiments/freeze_sweep/refugelike_freeze60",
|
|
"reps": 10,
|
|
"tower_overrides": {
|
|
"img": { "args": { "backbone": "refugelike", "freeze_ratio": 0.6 } }
|
|
}
|
|
},
|
|
|
|
{
|
|
"_note": "Fully frozen backbone (freeze_ratio=1.0) — image features are entirely fixed at the REFUGE-pretrained state, only the L1 bridge and downstream heads adapt. Provides the extreme end of the freeze-ratio sweep for the methods ablation table.",
|
|
"run_name": "experiments/freeze_sweep/refugelike_freeze100",
|
|
"reps": 10,
|
|
"tower_overrides": {
|
|
"img": { "args": { "backbone": "refugelike", "freeze_ratio": 1.0 } }
|
|
}
|
|
}
|
|
]
|